Test the verdict your routing runs on, every time it changes.
Your ICP model was checked once. Someone scored last year’s closed-won deals, most of them landed in Tier 1, and the model went live.
Since then the prompt has been edited, the model underneath it has been upgraded, and the accounts coming in look nothing like the ones you closed. Every one of those changes can move a verdict, and nothing in your CRM reports it. Tier 1 keeps filling up and the dashboard looks the same.
We ran our own golden set twice, with the same prompt, the same model and the same accounts. Some verdicts changed between the runs. A one-time backtest cannot see that, because it only runs once.
This week’s playbook is the workflow that catches it. Two checks before a verdict is allowed to route, and one readout after.
The workflow, in seven steps
- Write the ICP as one verdict field
- Label a golden set blind
- An agent returns one sourced verdict
- Check agreement against the labels
- Check that the verdict holds still
- Route by tier, with unknown as its own queue
- Read outcomes by the tier that routed them
Each step on the playbook page carries the reasoning, the failure it prevents, and a watch-for line from running it.
Two things the playbook covers that most ICP models skip
The fourth verdict. What happens to an account your data cannot decide, and why forcing it into a tier lets a guess route like a fact.
When to widen. How to read meetings and wins by the tier that routed each account, and the rule you write before the first readout that tells you when to give Tier 2 the Tier 1 treatment, or send the definition back for a rewrite.
What is in the implementation kit
Six markdown files and the playbook PDF, free by email from the playbook page.
- The fit definition, with four verdicts and room for a declared and an evidenced profile
- The blind labelling guide for the golden set
- The verdict prompt, with the check that runs before your CRM reads it
- The agreement and stability check, with the run record
- The monthly outcomes readout and the widen or pull back rule
- Ten sample companies and the answer key
Start with the sample set. Run the verdict prompt on it three times. If your verdicts match the key, including the two accounts that flip between runs, the setup is working.
Get the playbook and the kit →
A fit score tested once is a guess with a good memory.